Facial Action Unit-Based Deepfake Video Detection Using Deep Learning
Qasim Jaleel, Israa Hadi
Abstract
Deepfake videos are becoming more realistic, making them a menace. As a result of the development of deep learning techniques such as Generative adversarial networks (GAN), Deepfake has become closer to the truth. Widespread use of falsified videos and images on social media requires accurate detection. An identity switch (DeepFake) and an expression swap create facial modifications. This paper can detect deepfakes that are perfectly created. Traditional detection approaches that observe artifacts and pixel irregularities cannot keep up with modern technology. The paper is divided into two stages. In the first stage, the paper extracts facial action units from a person and creates a profile for him. This profile represents the behavior of his facial expressions, which differ from one person to another. This was done by building a deep learning network and training it based on a dataset. The second stage is testing, which involves taking videos, extracting facial action units, and testing them on the network to classify them as fake or real. The network has proven its ability to classify with high accuracy of %95.75 compared to traditional methods.